A finite state Markov quantizer for speech coding

A low-bit-rate codec based on an ergodic hidden Markov model (EHMM) is presented. A 256-state autoregressive Gaussian EHMM is trained on speech uttered by eight different speakers by means of the Baum-Welch algorithm. Initial estimates are obtained from vector quantization. The resulting EHMM is then utilized for a Viterbi decoding of incoming speech data. The state sequence obtained is frame synchronously encoded. The bit rate is gradually lowered by cutting off low-probability transitions and thus reducing the destination state encoding bit allocation requirements. The encoded spectra sequence is used, on the receiver side of the codec, for linear predictive coding synthesis. Global entropy and distortion measures for different bit rates are reported and compared to vector quantization results. Informal listening tests were performed by comparing the results of the proposed method at various bit rates and vector quantization using the same material